# PRE-REGISTRATION — the dense/MoE stopwatch (written before any scored call)

Registered 2026-08-19 ~18:10 UTC, ahead of a workday run beside live traffic
(operator-blessed: "run most of these benches now, today, during the workday —
rulesage-live traffic is low"). Contention from live requests is possible and
will be disclosed, per the august-arrivals precedent. The one leg this run
cannot lawfully take — the dense flagship with its draft head OFF (a runner
restart) — rides tonight's declared quiet window instead, alongside the
draft-head matrix whose draft-0 block doubles as that baseline.

## Cast (architecture from the runtime's own metadata, fetched 2026-08-19)

| tag | kind | total params | experts | quant |
|---|---|---|---|---|
| qwen3.8:27b | DENSE | 27.3B | — | Q4_K_M |
| gemma4:12b | DENSE | 11.9B | — | Q4_K_M |
| gemma4:26b | MoE | 25.8B | 8 used of 128 | Q4_K_M |
| nemotron-3.5-lightning:30b-a3b | MoE (hybrid) | 32.9B | 128 (a3b ≈ 3B active by tag) | Q4_K_M |

Three are standing residents; the a3b loads as the fifth model (the count cap
is 5; the fourth-model non-eviction receipt was taken this morning). Resident
calls pass keep_alive -1 explicitly; the a3b is released at run end. qwen3.8
is measured AS SERVED (draft head on, draft 4) and the rows say so.

## Rule

The C5 posture, unchanged: temperature 0, seed 0, think false, num_ctx 32768,
num_predict 256; prompt tiers ~1,000 and ~8,000 tokens (repetition-built, same
builder as sixteen's runs); n=10 per tier per model, serial calls, one warm-up
call per model per tier unscored. Metrics per call from the runtime's own
counters: prompt_eval_count/duration (prefill), eval_count/duration (decode),
load_duration, total. Resident bytes via /api/ps at run end.

## Predictions (falsifiable, written before the first scored call)

- **P1 (the headline bet):** gemma4:26b (MoE, 25.8B total) decodes FASTER than
  gemma4:12b (dense, 11.9B) — the payroll is bigger, the meeting is smaller.
- **P2:** the a3b decodes fastest of all four.
- **P3:** qwen3.8:27b as-served (draft-assisted) lands between the MoEs and
  a naive dense expectation; its draft-off truth (tonight's window) comes in
  slowest of the four.
- **P4:** prefill (tok/s) spreads across the cast far less than decode does —
  the group photo is parallel for everyone.
- **P5:** resident VRAM tracks TOTAL quantized size, not active size — the
  whole payroll must be in the building.

A prediction that fails is published as failed. No threshold table exists to
protect; this is an explainer's evidence, not a seat exam.

## Artifacts

runs.json (every call, every counter) + RESULTS.md (medians per cell, the
prediction verdicts, contention disclosure) in this directory. Box-neutral
vocabulary throughout: "the inference workstation." Run id:
moe-dense-stopwatch-2026-08-19.